xref: /linux/net/sched/sch_generic.c (revision 6ebe6dbd6886af07b102aca42e44edbee94a22d9)
1 /*
2  * net/sched/sch_generic.c	Generic packet scheduler routines.
3  *
4  *		This program is free software; you can redistribute it and/or
5  *		modify it under the terms of the GNU General Public License
6  *		as published by the Free Software Foundation; either version
7  *		2 of the License, or (at your option) any later version.
8  *
9  * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11  *              - Ingress support
12  */
13 
14 #include <linux/bitops.h>
15 #include <linux/module.h>
16 #include <linux/types.h>
17 #include <linux/kernel.h>
18 #include <linux/sched.h>
19 #include <linux/string.h>
20 #include <linux/errno.h>
21 #include <linux/netdevice.h>
22 #include <linux/skbuff.h>
23 #include <linux/rtnetlink.h>
24 #include <linux/init.h>
25 #include <linux/rcupdate.h>
26 #include <linux/list.h>
27 #include <linux/slab.h>
28 #include <linux/if_vlan.h>
29 #include <linux/skb_array.h>
30 #include <linux/if_macvlan.h>
31 #include <net/sch_generic.h>
32 #include <net/pkt_sched.h>
33 #include <net/dst.h>
34 #include <trace/events/qdisc.h>
35 #include <net/xfrm.h>
36 
37 /* Qdisc to use by default */
38 const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
39 EXPORT_SYMBOL(default_qdisc_ops);
40 
41 /* Main transmission queue. */
42 
43 /* Modifications to data participating in scheduling must be protected with
44  * qdisc_lock(qdisc) spinlock.
45  *
46  * The idea is the following:
47  * - enqueue, dequeue are serialized via qdisc root lock
48  * - ingress filtering is also serialized via qdisc root lock
49  * - updates to tree and tree walking are only done under the rtnl mutex.
50  */
51 
52 static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
53 {
54 	const struct netdev_queue *txq = q->dev_queue;
55 	spinlock_t *lock = NULL;
56 	struct sk_buff *skb;
57 
58 	if (q->flags & TCQ_F_NOLOCK) {
59 		lock = qdisc_lock(q);
60 		spin_lock(lock);
61 	}
62 
63 	skb = skb_peek(&q->skb_bad_txq);
64 	if (skb) {
65 		/* check the reason of requeuing without tx lock first */
66 		txq = skb_get_tx_queue(txq->dev, skb);
67 		if (!netif_xmit_frozen_or_stopped(txq)) {
68 			skb = __skb_dequeue(&q->skb_bad_txq);
69 			if (qdisc_is_percpu_stats(q)) {
70 				qdisc_qstats_cpu_backlog_dec(q, skb);
71 				qdisc_qstats_cpu_qlen_dec(q);
72 			} else {
73 				qdisc_qstats_backlog_dec(q, skb);
74 				q->q.qlen--;
75 			}
76 		} else {
77 			skb = NULL;
78 		}
79 	}
80 
81 	if (lock)
82 		spin_unlock(lock);
83 
84 	return skb;
85 }
86 
87 static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
88 {
89 	struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
90 
91 	if (unlikely(skb))
92 		skb = __skb_dequeue_bad_txq(q);
93 
94 	return skb;
95 }
96 
97 static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
98 					     struct sk_buff *skb)
99 {
100 	spinlock_t *lock = NULL;
101 
102 	if (q->flags & TCQ_F_NOLOCK) {
103 		lock = qdisc_lock(q);
104 		spin_lock(lock);
105 	}
106 
107 	__skb_queue_tail(&q->skb_bad_txq, skb);
108 
109 	if (lock)
110 		spin_unlock(lock);
111 }
112 
113 static inline int __dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
114 {
115 	__skb_queue_head(&q->gso_skb, skb);
116 	q->qstats.requeues++;
117 	qdisc_qstats_backlog_inc(q, skb);
118 	q->q.qlen++;	/* it's still part of the queue */
119 	__netif_schedule(q);
120 
121 	return 0;
122 }
123 
124 static inline int dev_requeue_skb_locked(struct sk_buff *skb, struct Qdisc *q)
125 {
126 	spinlock_t *lock = qdisc_lock(q);
127 
128 	spin_lock(lock);
129 	__skb_queue_tail(&q->gso_skb, skb);
130 	spin_unlock(lock);
131 
132 	qdisc_qstats_cpu_requeues_inc(q);
133 	qdisc_qstats_cpu_backlog_inc(q, skb);
134 	qdisc_qstats_cpu_qlen_inc(q);
135 	__netif_schedule(q);
136 
137 	return 0;
138 }
139 
140 static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
141 {
142 	if (q->flags & TCQ_F_NOLOCK)
143 		return dev_requeue_skb_locked(skb, q);
144 	else
145 		return __dev_requeue_skb(skb, q);
146 }
147 
148 static void try_bulk_dequeue_skb(struct Qdisc *q,
149 				 struct sk_buff *skb,
150 				 const struct netdev_queue *txq,
151 				 int *packets)
152 {
153 	int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
154 
155 	while (bytelimit > 0) {
156 		struct sk_buff *nskb = q->dequeue(q);
157 
158 		if (!nskb)
159 			break;
160 
161 		bytelimit -= nskb->len; /* covers GSO len */
162 		skb->next = nskb;
163 		skb = nskb;
164 		(*packets)++; /* GSO counts as one pkt */
165 	}
166 	skb->next = NULL;
167 }
168 
169 /* This variant of try_bulk_dequeue_skb() makes sure
170  * all skbs in the chain are for the same txq
171  */
172 static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
173 				      struct sk_buff *skb,
174 				      int *packets)
175 {
176 	int mapping = skb_get_queue_mapping(skb);
177 	struct sk_buff *nskb;
178 	int cnt = 0;
179 
180 	do {
181 		nskb = q->dequeue(q);
182 		if (!nskb)
183 			break;
184 		if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
185 			qdisc_enqueue_skb_bad_txq(q, nskb);
186 
187 			if (qdisc_is_percpu_stats(q)) {
188 				qdisc_qstats_cpu_backlog_inc(q, nskb);
189 				qdisc_qstats_cpu_qlen_inc(q);
190 			} else {
191 				qdisc_qstats_backlog_inc(q, nskb);
192 				q->q.qlen++;
193 			}
194 			break;
195 		}
196 		skb->next = nskb;
197 		skb = nskb;
198 	} while (++cnt < 8);
199 	(*packets) += cnt;
200 	skb->next = NULL;
201 }
202 
203 /* Note that dequeue_skb can possibly return a SKB list (via skb->next).
204  * A requeued skb (via q->gso_skb) can also be a SKB list.
205  */
206 static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
207 				   int *packets)
208 {
209 	const struct netdev_queue *txq = q->dev_queue;
210 	struct sk_buff *skb = NULL;
211 
212 	*packets = 1;
213 	if (unlikely(!skb_queue_empty(&q->gso_skb))) {
214 		spinlock_t *lock = NULL;
215 
216 		if (q->flags & TCQ_F_NOLOCK) {
217 			lock = qdisc_lock(q);
218 			spin_lock(lock);
219 		}
220 
221 		skb = skb_peek(&q->gso_skb);
222 
223 		/* skb may be null if another cpu pulls gso_skb off in between
224 		 * empty check and lock.
225 		 */
226 		if (!skb) {
227 			if (lock)
228 				spin_unlock(lock);
229 			goto validate;
230 		}
231 
232 		/* skb in gso_skb were already validated */
233 		*validate = false;
234 		if (xfrm_offload(skb))
235 			*validate = true;
236 		/* check the reason of requeuing without tx lock first */
237 		txq = skb_get_tx_queue(txq->dev, skb);
238 		if (!netif_xmit_frozen_or_stopped(txq)) {
239 			skb = __skb_dequeue(&q->gso_skb);
240 			if (qdisc_is_percpu_stats(q)) {
241 				qdisc_qstats_cpu_backlog_dec(q, skb);
242 				qdisc_qstats_cpu_qlen_dec(q);
243 			} else {
244 				qdisc_qstats_backlog_dec(q, skb);
245 				q->q.qlen--;
246 			}
247 		} else {
248 			skb = NULL;
249 		}
250 		if (lock)
251 			spin_unlock(lock);
252 		goto trace;
253 	}
254 validate:
255 	*validate = true;
256 
257 	if ((q->flags & TCQ_F_ONETXQUEUE) &&
258 	    netif_xmit_frozen_or_stopped(txq))
259 		return skb;
260 
261 	skb = qdisc_dequeue_skb_bad_txq(q);
262 	if (unlikely(skb))
263 		goto bulk;
264 	skb = q->dequeue(q);
265 	if (skb) {
266 bulk:
267 		if (qdisc_may_bulk(q))
268 			try_bulk_dequeue_skb(q, skb, txq, packets);
269 		else
270 			try_bulk_dequeue_skb_slow(q, skb, packets);
271 	}
272 trace:
273 	trace_qdisc_dequeue(q, txq, *packets, skb);
274 	return skb;
275 }
276 
277 /*
278  * Transmit possibly several skbs, and handle the return status as
279  * required. Owning running seqcount bit guarantees that
280  * only one CPU can execute this function.
281  *
282  * Returns to the caller:
283  *				false  - hardware queue frozen backoff
284  *				true   - feel free to send more pkts
285  */
286 bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
287 		     struct net_device *dev, struct netdev_queue *txq,
288 		     spinlock_t *root_lock, bool validate)
289 {
290 	int ret = NETDEV_TX_BUSY;
291 	bool again = false;
292 
293 	/* And release qdisc */
294 	if (root_lock)
295 		spin_unlock(root_lock);
296 
297 	/* Note that we validate skb (GSO, checksum, ...) outside of locks */
298 	if (validate)
299 		skb = validate_xmit_skb_list(skb, dev, &again);
300 
301 #ifdef CONFIG_XFRM_OFFLOAD
302 	if (unlikely(again)) {
303 		if (root_lock)
304 			spin_lock(root_lock);
305 
306 		dev_requeue_skb(skb, q);
307 		return false;
308 	}
309 #endif
310 
311 	if (likely(skb)) {
312 		HARD_TX_LOCK(dev, txq, smp_processor_id());
313 		if (!netif_xmit_frozen_or_stopped(txq))
314 			skb = dev_hard_start_xmit(skb, dev, txq, &ret);
315 
316 		HARD_TX_UNLOCK(dev, txq);
317 	} else {
318 		if (root_lock)
319 			spin_lock(root_lock);
320 		return true;
321 	}
322 
323 	if (root_lock)
324 		spin_lock(root_lock);
325 
326 	if (!dev_xmit_complete(ret)) {
327 		/* Driver returned NETDEV_TX_BUSY - requeue skb */
328 		if (unlikely(ret != NETDEV_TX_BUSY))
329 			net_warn_ratelimited("BUG %s code %d qlen %d\n",
330 					     dev->name, ret, q->q.qlen);
331 
332 		dev_requeue_skb(skb, q);
333 		return false;
334 	}
335 
336 	if (ret && netif_xmit_frozen_or_stopped(txq))
337 		return false;
338 
339 	return true;
340 }
341 
342 /*
343  * NOTE: Called under qdisc_lock(q) with locally disabled BH.
344  *
345  * running seqcount guarantees only one CPU can process
346  * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
347  * this queue.
348  *
349  *  netif_tx_lock serializes accesses to device driver.
350  *
351  *  qdisc_lock(q) and netif_tx_lock are mutually exclusive,
352  *  if one is grabbed, another must be free.
353  *
354  * Note, that this procedure can be called by a watchdog timer
355  *
356  * Returns to the caller:
357  *				0  - queue is empty or throttled.
358  *				>0 - queue is not empty.
359  *
360  */
361 static inline bool qdisc_restart(struct Qdisc *q, int *packets)
362 {
363 	spinlock_t *root_lock = NULL;
364 	struct netdev_queue *txq;
365 	struct net_device *dev;
366 	struct sk_buff *skb;
367 	bool validate;
368 
369 	/* Dequeue packet */
370 	skb = dequeue_skb(q, &validate, packets);
371 	if (unlikely(!skb))
372 		return false;
373 
374 	if (!(q->flags & TCQ_F_NOLOCK))
375 		root_lock = qdisc_lock(q);
376 
377 	dev = qdisc_dev(q);
378 	txq = skb_get_tx_queue(dev, skb);
379 
380 	return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
381 }
382 
383 void __qdisc_run(struct Qdisc *q)
384 {
385 	int quota = dev_tx_weight;
386 	int packets;
387 
388 	while (qdisc_restart(q, &packets)) {
389 		/*
390 		 * Ordered by possible occurrence: Postpone processing if
391 		 * 1. we've exceeded packet quota
392 		 * 2. another process needs the CPU;
393 		 */
394 		quota -= packets;
395 		if (quota <= 0 || need_resched()) {
396 			__netif_schedule(q);
397 			break;
398 		}
399 	}
400 }
401 
402 unsigned long dev_trans_start(struct net_device *dev)
403 {
404 	unsigned long val, res;
405 	unsigned int i;
406 
407 	if (is_vlan_dev(dev))
408 		dev = vlan_dev_real_dev(dev);
409 	else if (netif_is_macvlan(dev))
410 		dev = macvlan_dev_real_dev(dev);
411 	res = netdev_get_tx_queue(dev, 0)->trans_start;
412 	for (i = 1; i < dev->num_tx_queues; i++) {
413 		val = netdev_get_tx_queue(dev, i)->trans_start;
414 		if (val && time_after(val, res))
415 			res = val;
416 	}
417 
418 	return res;
419 }
420 EXPORT_SYMBOL(dev_trans_start);
421 
422 static void dev_watchdog(struct timer_list *t)
423 {
424 	struct net_device *dev = from_timer(dev, t, watchdog_timer);
425 
426 	netif_tx_lock(dev);
427 	if (!qdisc_tx_is_noop(dev)) {
428 		if (netif_device_present(dev) &&
429 		    netif_running(dev) &&
430 		    netif_carrier_ok(dev)) {
431 			int some_queue_timedout = 0;
432 			unsigned int i;
433 			unsigned long trans_start;
434 
435 			for (i = 0; i < dev->num_tx_queues; i++) {
436 				struct netdev_queue *txq;
437 
438 				txq = netdev_get_tx_queue(dev, i);
439 				trans_start = txq->trans_start;
440 				if (netif_xmit_stopped(txq) &&
441 				    time_after(jiffies, (trans_start +
442 							 dev->watchdog_timeo))) {
443 					some_queue_timedout = 1;
444 					txq->trans_timeout++;
445 					break;
446 				}
447 			}
448 
449 			if (some_queue_timedout) {
450 				WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
451 				       dev->name, netdev_drivername(dev), i);
452 				dev->netdev_ops->ndo_tx_timeout(dev);
453 			}
454 			if (!mod_timer(&dev->watchdog_timer,
455 				       round_jiffies(jiffies +
456 						     dev->watchdog_timeo)))
457 				dev_hold(dev);
458 		}
459 	}
460 	netif_tx_unlock(dev);
461 
462 	dev_put(dev);
463 }
464 
465 void __netdev_watchdog_up(struct net_device *dev)
466 {
467 	if (dev->netdev_ops->ndo_tx_timeout) {
468 		if (dev->watchdog_timeo <= 0)
469 			dev->watchdog_timeo = 5*HZ;
470 		if (!mod_timer(&dev->watchdog_timer,
471 			       round_jiffies(jiffies + dev->watchdog_timeo)))
472 			dev_hold(dev);
473 	}
474 }
475 
476 static void dev_watchdog_up(struct net_device *dev)
477 {
478 	__netdev_watchdog_up(dev);
479 }
480 
481 static void dev_watchdog_down(struct net_device *dev)
482 {
483 	netif_tx_lock_bh(dev);
484 	if (del_timer(&dev->watchdog_timer))
485 		dev_put(dev);
486 	netif_tx_unlock_bh(dev);
487 }
488 
489 /**
490  *	netif_carrier_on - set carrier
491  *	@dev: network device
492  *
493  * Device has detected that carrier.
494  */
495 void netif_carrier_on(struct net_device *dev)
496 {
497 	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
498 		if (dev->reg_state == NETREG_UNINITIALIZED)
499 			return;
500 		atomic_inc(&dev->carrier_changes);
501 		linkwatch_fire_event(dev);
502 		if (netif_running(dev))
503 			__netdev_watchdog_up(dev);
504 	}
505 }
506 EXPORT_SYMBOL(netif_carrier_on);
507 
508 /**
509  *	netif_carrier_off - clear carrier
510  *	@dev: network device
511  *
512  * Device has detected loss of carrier.
513  */
514 void netif_carrier_off(struct net_device *dev)
515 {
516 	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
517 		if (dev->reg_state == NETREG_UNINITIALIZED)
518 			return;
519 		atomic_inc(&dev->carrier_changes);
520 		linkwatch_fire_event(dev);
521 	}
522 }
523 EXPORT_SYMBOL(netif_carrier_off);
524 
525 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
526    under all circumstances. It is difficult to invent anything faster or
527    cheaper.
528  */
529 
530 static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
531 			struct sk_buff **to_free)
532 {
533 	__qdisc_drop(skb, to_free);
534 	return NET_XMIT_CN;
535 }
536 
537 static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
538 {
539 	return NULL;
540 }
541 
542 struct Qdisc_ops noop_qdisc_ops __read_mostly = {
543 	.id		=	"noop",
544 	.priv_size	=	0,
545 	.enqueue	=	noop_enqueue,
546 	.dequeue	=	noop_dequeue,
547 	.peek		=	noop_dequeue,
548 	.owner		=	THIS_MODULE,
549 };
550 
551 static struct netdev_queue noop_netdev_queue = {
552 	.qdisc		=	&noop_qdisc,
553 	.qdisc_sleeping	=	&noop_qdisc,
554 };
555 
556 struct Qdisc noop_qdisc = {
557 	.enqueue	=	noop_enqueue,
558 	.dequeue	=	noop_dequeue,
559 	.flags		=	TCQ_F_BUILTIN,
560 	.ops		=	&noop_qdisc_ops,
561 	.q.lock		=	__SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
562 	.dev_queue	=	&noop_netdev_queue,
563 	.running	=	SEQCNT_ZERO(noop_qdisc.running),
564 	.busylock	=	__SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
565 };
566 EXPORT_SYMBOL(noop_qdisc);
567 
568 static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
569 			struct netlink_ext_ack *extack)
570 {
571 	/* register_qdisc() assigns a default of noop_enqueue if unset,
572 	 * but __dev_queue_xmit() treats noqueue only as such
573 	 * if this is NULL - so clear it here. */
574 	qdisc->enqueue = NULL;
575 	return 0;
576 }
577 
578 struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
579 	.id		=	"noqueue",
580 	.priv_size	=	0,
581 	.init		=	noqueue_init,
582 	.enqueue	=	noop_enqueue,
583 	.dequeue	=	noop_dequeue,
584 	.peek		=	noop_dequeue,
585 	.owner		=	THIS_MODULE,
586 };
587 
588 static const u8 prio2band[TC_PRIO_MAX + 1] = {
589 	1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
590 };
591 
592 /* 3-band FIFO queue: old style, but should be a bit faster than
593    generic prio+fifo combination.
594  */
595 
596 #define PFIFO_FAST_BANDS 3
597 
598 /*
599  * Private data for a pfifo_fast scheduler containing:
600  *	- rings for priority bands
601  */
602 struct pfifo_fast_priv {
603 	struct skb_array q[PFIFO_FAST_BANDS];
604 };
605 
606 static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
607 					  int band)
608 {
609 	return &priv->q[band];
610 }
611 
612 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
613 			      struct sk_buff **to_free)
614 {
615 	int band = prio2band[skb->priority & TC_PRIO_MAX];
616 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
617 	struct skb_array *q = band2list(priv, band);
618 	int err;
619 
620 	err = skb_array_produce(q, skb);
621 
622 	if (unlikely(err))
623 		return qdisc_drop_cpu(skb, qdisc, to_free);
624 
625 	qdisc_qstats_cpu_qlen_inc(qdisc);
626 	qdisc_qstats_cpu_backlog_inc(qdisc, skb);
627 	return NET_XMIT_SUCCESS;
628 }
629 
630 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
631 {
632 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
633 	struct sk_buff *skb = NULL;
634 	int band;
635 
636 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
637 		struct skb_array *q = band2list(priv, band);
638 
639 		if (__skb_array_empty(q))
640 			continue;
641 
642 		skb = skb_array_consume_bh(q);
643 	}
644 	if (likely(skb)) {
645 		qdisc_qstats_cpu_backlog_dec(qdisc, skb);
646 		qdisc_bstats_cpu_update(qdisc, skb);
647 		qdisc_qstats_cpu_qlen_dec(qdisc);
648 	}
649 
650 	return skb;
651 }
652 
653 static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
654 {
655 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
656 	struct sk_buff *skb = NULL;
657 	int band;
658 
659 	for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
660 		struct skb_array *q = band2list(priv, band);
661 
662 		skb = __skb_array_peek(q);
663 	}
664 
665 	return skb;
666 }
667 
668 static void pfifo_fast_reset(struct Qdisc *qdisc)
669 {
670 	int i, band;
671 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
672 
673 	for (band = 0; band < PFIFO_FAST_BANDS; band++) {
674 		struct skb_array *q = band2list(priv, band);
675 		struct sk_buff *skb;
676 
677 		/* NULL ring is possible if destroy path is due to a failed
678 		 * skb_array_init() in pfifo_fast_init() case.
679 		 */
680 		if (!q->ring.queue)
681 			continue;
682 
683 		while ((skb = skb_array_consume_bh(q)) != NULL)
684 			kfree_skb(skb);
685 	}
686 
687 	for_each_possible_cpu(i) {
688 		struct gnet_stats_queue *q = per_cpu_ptr(qdisc->cpu_qstats, i);
689 
690 		q->backlog = 0;
691 		q->qlen = 0;
692 	}
693 }
694 
695 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
696 {
697 	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
698 
699 	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
700 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
701 		goto nla_put_failure;
702 	return skb->len;
703 
704 nla_put_failure:
705 	return -1;
706 }
707 
708 static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
709 			   struct netlink_ext_ack *extack)
710 {
711 	unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
712 	struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
713 	int prio;
714 
715 	/* guard against zero length rings */
716 	if (!qlen)
717 		return -EINVAL;
718 
719 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
720 		struct skb_array *q = band2list(priv, prio);
721 		int err;
722 
723 		err = skb_array_init(q, qlen, GFP_KERNEL);
724 		if (err)
725 			return -ENOMEM;
726 	}
727 
728 	/* Can by-pass the queue discipline */
729 	qdisc->flags |= TCQ_F_CAN_BYPASS;
730 	return 0;
731 }
732 
733 static void pfifo_fast_destroy(struct Qdisc *sch)
734 {
735 	struct pfifo_fast_priv *priv = qdisc_priv(sch);
736 	int prio;
737 
738 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
739 		struct skb_array *q = band2list(priv, prio);
740 
741 		/* NULL ring is possible if destroy path is due to a failed
742 		 * skb_array_init() in pfifo_fast_init() case.
743 		 */
744 		if (!q->ring.queue)
745 			continue;
746 		/* Destroy ring but no need to kfree_skb because a call to
747 		 * pfifo_fast_reset() has already done that work.
748 		 */
749 		ptr_ring_cleanup(&q->ring, NULL);
750 	}
751 }
752 
753 struct Qdisc_ops pfifo_fast_ops __read_mostly = {
754 	.id		=	"pfifo_fast",
755 	.priv_size	=	sizeof(struct pfifo_fast_priv),
756 	.enqueue	=	pfifo_fast_enqueue,
757 	.dequeue	=	pfifo_fast_dequeue,
758 	.peek		=	pfifo_fast_peek,
759 	.init		=	pfifo_fast_init,
760 	.destroy	=	pfifo_fast_destroy,
761 	.reset		=	pfifo_fast_reset,
762 	.dump		=	pfifo_fast_dump,
763 	.owner		=	THIS_MODULE,
764 	.static_flags	=	TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
765 };
766 EXPORT_SYMBOL(pfifo_fast_ops);
767 
768 static struct lock_class_key qdisc_tx_busylock;
769 static struct lock_class_key qdisc_running_key;
770 
771 struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
772 			  const struct Qdisc_ops *ops,
773 			  struct netlink_ext_ack *extack)
774 {
775 	void *p;
776 	struct Qdisc *sch;
777 	unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
778 	int err = -ENOBUFS;
779 	struct net_device *dev;
780 
781 	if (!dev_queue) {
782 		NL_SET_ERR_MSG(extack, "No device queue given");
783 		err = -EINVAL;
784 		goto errout;
785 	}
786 
787 	dev = dev_queue->dev;
788 	p = kzalloc_node(size, GFP_KERNEL,
789 			 netdev_queue_numa_node_read(dev_queue));
790 
791 	if (!p)
792 		goto errout;
793 	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
794 	/* if we got non aligned memory, ask more and do alignment ourself */
795 	if (sch != p) {
796 		kfree(p);
797 		p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
798 				 netdev_queue_numa_node_read(dev_queue));
799 		if (!p)
800 			goto errout;
801 		sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
802 		sch->padded = (char *) sch - (char *) p;
803 	}
804 	__skb_queue_head_init(&sch->gso_skb);
805 	__skb_queue_head_init(&sch->skb_bad_txq);
806 	qdisc_skb_head_init(&sch->q);
807 	spin_lock_init(&sch->q.lock);
808 
809 	if (ops->static_flags & TCQ_F_CPUSTATS) {
810 		sch->cpu_bstats =
811 			netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
812 		if (!sch->cpu_bstats)
813 			goto errout1;
814 
815 		sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
816 		if (!sch->cpu_qstats) {
817 			free_percpu(sch->cpu_bstats);
818 			goto errout1;
819 		}
820 	}
821 
822 	spin_lock_init(&sch->busylock);
823 	lockdep_set_class(&sch->busylock,
824 			  dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
825 
826 	seqcount_init(&sch->running);
827 	lockdep_set_class(&sch->running,
828 			  dev->qdisc_running_key ?: &qdisc_running_key);
829 
830 	sch->ops = ops;
831 	sch->flags = ops->static_flags;
832 	sch->enqueue = ops->enqueue;
833 	sch->dequeue = ops->dequeue;
834 	sch->dev_queue = dev_queue;
835 	dev_hold(dev);
836 	refcount_set(&sch->refcnt, 1);
837 
838 	return sch;
839 errout1:
840 	kfree(p);
841 errout:
842 	return ERR_PTR(err);
843 }
844 
845 struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
846 				const struct Qdisc_ops *ops,
847 				unsigned int parentid,
848 				struct netlink_ext_ack *extack)
849 {
850 	struct Qdisc *sch;
851 
852 	if (!try_module_get(ops->owner)) {
853 		NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
854 		return NULL;
855 	}
856 
857 	sch = qdisc_alloc(dev_queue, ops, extack);
858 	if (IS_ERR(sch)) {
859 		module_put(ops->owner);
860 		return NULL;
861 	}
862 	sch->parent = parentid;
863 
864 	if (!ops->init || ops->init(sch, NULL, extack) == 0)
865 		return sch;
866 
867 	qdisc_destroy(sch);
868 	return NULL;
869 }
870 EXPORT_SYMBOL(qdisc_create_dflt);
871 
872 /* Under qdisc_lock(qdisc) and BH! */
873 
874 void qdisc_reset(struct Qdisc *qdisc)
875 {
876 	const struct Qdisc_ops *ops = qdisc->ops;
877 	struct sk_buff *skb, *tmp;
878 
879 	if (ops->reset)
880 		ops->reset(qdisc);
881 
882 	skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
883 		__skb_unlink(skb, &qdisc->gso_skb);
884 		kfree_skb_list(skb);
885 	}
886 
887 	skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
888 		__skb_unlink(skb, &qdisc->skb_bad_txq);
889 		kfree_skb_list(skb);
890 	}
891 
892 	qdisc->q.qlen = 0;
893 	qdisc->qstats.backlog = 0;
894 }
895 EXPORT_SYMBOL(qdisc_reset);
896 
897 static void qdisc_free(struct Qdisc *qdisc)
898 {
899 	if (qdisc_is_percpu_stats(qdisc)) {
900 		free_percpu(qdisc->cpu_bstats);
901 		free_percpu(qdisc->cpu_qstats);
902 	}
903 
904 	kfree((char *) qdisc - qdisc->padded);
905 }
906 
907 void qdisc_destroy(struct Qdisc *qdisc)
908 {
909 	const struct Qdisc_ops  *ops = qdisc->ops;
910 	struct sk_buff *skb, *tmp;
911 
912 	if (qdisc->flags & TCQ_F_BUILTIN ||
913 	    !refcount_dec_and_test(&qdisc->refcnt))
914 		return;
915 
916 #ifdef CONFIG_NET_SCHED
917 	qdisc_hash_del(qdisc);
918 
919 	qdisc_put_stab(rtnl_dereference(qdisc->stab));
920 #endif
921 	gen_kill_estimator(&qdisc->rate_est);
922 	if (ops->reset)
923 		ops->reset(qdisc);
924 	if (ops->destroy)
925 		ops->destroy(qdisc);
926 
927 	module_put(ops->owner);
928 	dev_put(qdisc_dev(qdisc));
929 
930 	skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
931 		__skb_unlink(skb, &qdisc->gso_skb);
932 		kfree_skb_list(skb);
933 	}
934 
935 	skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
936 		__skb_unlink(skb, &qdisc->skb_bad_txq);
937 		kfree_skb_list(skb);
938 	}
939 
940 	qdisc_free(qdisc);
941 }
942 EXPORT_SYMBOL(qdisc_destroy);
943 
944 /* Attach toplevel qdisc to device queue. */
945 struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
946 			      struct Qdisc *qdisc)
947 {
948 	struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
949 	spinlock_t *root_lock;
950 
951 	root_lock = qdisc_lock(oqdisc);
952 	spin_lock_bh(root_lock);
953 
954 	/* ... and graft new one */
955 	if (qdisc == NULL)
956 		qdisc = &noop_qdisc;
957 	dev_queue->qdisc_sleeping = qdisc;
958 	rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
959 
960 	spin_unlock_bh(root_lock);
961 
962 	return oqdisc;
963 }
964 EXPORT_SYMBOL(dev_graft_qdisc);
965 
966 static void attach_one_default_qdisc(struct net_device *dev,
967 				     struct netdev_queue *dev_queue,
968 				     void *_unused)
969 {
970 	struct Qdisc *qdisc;
971 	const struct Qdisc_ops *ops = default_qdisc_ops;
972 
973 	if (dev->priv_flags & IFF_NO_QUEUE)
974 		ops = &noqueue_qdisc_ops;
975 
976 	qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
977 	if (!qdisc) {
978 		netdev_info(dev, "activation failed\n");
979 		return;
980 	}
981 	if (!netif_is_multiqueue(dev))
982 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
983 	dev_queue->qdisc_sleeping = qdisc;
984 }
985 
986 static void attach_default_qdiscs(struct net_device *dev)
987 {
988 	struct netdev_queue *txq;
989 	struct Qdisc *qdisc;
990 
991 	txq = netdev_get_tx_queue(dev, 0);
992 
993 	if (!netif_is_multiqueue(dev) ||
994 	    dev->priv_flags & IFF_NO_QUEUE) {
995 		netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
996 		dev->qdisc = txq->qdisc_sleeping;
997 		qdisc_refcount_inc(dev->qdisc);
998 	} else {
999 		qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1000 		if (qdisc) {
1001 			dev->qdisc = qdisc;
1002 			qdisc->ops->attach(qdisc);
1003 		}
1004 	}
1005 #ifdef CONFIG_NET_SCHED
1006 	if (dev->qdisc != &noop_qdisc)
1007 		qdisc_hash_add(dev->qdisc, false);
1008 #endif
1009 }
1010 
1011 static void transition_one_qdisc(struct net_device *dev,
1012 				 struct netdev_queue *dev_queue,
1013 				 void *_need_watchdog)
1014 {
1015 	struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1016 	int *need_watchdog_p = _need_watchdog;
1017 
1018 	if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1019 		clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1020 
1021 	rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1022 	if (need_watchdog_p) {
1023 		dev_queue->trans_start = 0;
1024 		*need_watchdog_p = 1;
1025 	}
1026 }
1027 
1028 void dev_activate(struct net_device *dev)
1029 {
1030 	int need_watchdog;
1031 
1032 	/* No queueing discipline is attached to device;
1033 	 * create default one for devices, which need queueing
1034 	 * and noqueue_qdisc for virtual interfaces
1035 	 */
1036 
1037 	if (dev->qdisc == &noop_qdisc)
1038 		attach_default_qdiscs(dev);
1039 
1040 	if (!netif_carrier_ok(dev))
1041 		/* Delay activation until next carrier-on event */
1042 		return;
1043 
1044 	need_watchdog = 0;
1045 	netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1046 	if (dev_ingress_queue(dev))
1047 		transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1048 
1049 	if (need_watchdog) {
1050 		netif_trans_update(dev);
1051 		dev_watchdog_up(dev);
1052 	}
1053 }
1054 EXPORT_SYMBOL(dev_activate);
1055 
1056 static void dev_deactivate_queue(struct net_device *dev,
1057 				 struct netdev_queue *dev_queue,
1058 				 void *_qdisc_default)
1059 {
1060 	struct Qdisc *qdisc_default = _qdisc_default;
1061 	struct Qdisc *qdisc;
1062 
1063 	qdisc = rtnl_dereference(dev_queue->qdisc);
1064 	if (qdisc) {
1065 		spin_lock_bh(qdisc_lock(qdisc));
1066 
1067 		if (!(qdisc->flags & TCQ_F_BUILTIN))
1068 			set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1069 
1070 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1071 		qdisc_reset(qdisc);
1072 
1073 		spin_unlock_bh(qdisc_lock(qdisc));
1074 	}
1075 }
1076 
1077 static bool some_qdisc_is_busy(struct net_device *dev)
1078 {
1079 	unsigned int i;
1080 
1081 	for (i = 0; i < dev->num_tx_queues; i++) {
1082 		struct netdev_queue *dev_queue;
1083 		spinlock_t *root_lock;
1084 		struct Qdisc *q;
1085 		int val;
1086 
1087 		dev_queue = netdev_get_tx_queue(dev, i);
1088 		q = dev_queue->qdisc_sleeping;
1089 
1090 		if (q->flags & TCQ_F_NOLOCK) {
1091 			val = test_bit(__QDISC_STATE_SCHED, &q->state);
1092 		} else {
1093 			root_lock = qdisc_lock(q);
1094 			spin_lock_bh(root_lock);
1095 
1096 			val = (qdisc_is_running(q) ||
1097 			       test_bit(__QDISC_STATE_SCHED, &q->state));
1098 
1099 			spin_unlock_bh(root_lock);
1100 		}
1101 
1102 		if (val)
1103 			return true;
1104 	}
1105 	return false;
1106 }
1107 
1108 static void dev_qdisc_reset(struct net_device *dev,
1109 			    struct netdev_queue *dev_queue,
1110 			    void *none)
1111 {
1112 	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1113 
1114 	if (qdisc)
1115 		qdisc_reset(qdisc);
1116 }
1117 
1118 /**
1119  * 	dev_deactivate_many - deactivate transmissions on several devices
1120  * 	@head: list of devices to deactivate
1121  *
1122  *	This function returns only when all outstanding transmissions
1123  *	have completed, unless all devices are in dismantle phase.
1124  */
1125 void dev_deactivate_many(struct list_head *head)
1126 {
1127 	struct net_device *dev;
1128 
1129 	list_for_each_entry(dev, head, close_list) {
1130 		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1131 					 &noop_qdisc);
1132 		if (dev_ingress_queue(dev))
1133 			dev_deactivate_queue(dev, dev_ingress_queue(dev),
1134 					     &noop_qdisc);
1135 
1136 		dev_watchdog_down(dev);
1137 	}
1138 
1139 	/* Wait for outstanding qdisc-less dev_queue_xmit calls.
1140 	 * This is avoided if all devices are in dismantle phase :
1141 	 * Caller will call synchronize_net() for us
1142 	 */
1143 	synchronize_net();
1144 
1145 	/* Wait for outstanding qdisc_run calls. */
1146 	list_for_each_entry(dev, head, close_list) {
1147 		while (some_qdisc_is_busy(dev))
1148 			yield();
1149 		/* The new qdisc is assigned at this point so we can safely
1150 		 * unwind stale skb lists and qdisc statistics
1151 		 */
1152 		netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1153 		if (dev_ingress_queue(dev))
1154 			dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1155 	}
1156 }
1157 
1158 void dev_deactivate(struct net_device *dev)
1159 {
1160 	LIST_HEAD(single);
1161 
1162 	list_add(&dev->close_list, &single);
1163 	dev_deactivate_many(&single);
1164 	list_del(&single);
1165 }
1166 EXPORT_SYMBOL(dev_deactivate);
1167 
1168 static void dev_init_scheduler_queue(struct net_device *dev,
1169 				     struct netdev_queue *dev_queue,
1170 				     void *_qdisc)
1171 {
1172 	struct Qdisc *qdisc = _qdisc;
1173 
1174 	rcu_assign_pointer(dev_queue->qdisc, qdisc);
1175 	dev_queue->qdisc_sleeping = qdisc;
1176 	__skb_queue_head_init(&qdisc->gso_skb);
1177 	__skb_queue_head_init(&qdisc->skb_bad_txq);
1178 }
1179 
1180 void dev_init_scheduler(struct net_device *dev)
1181 {
1182 	dev->qdisc = &noop_qdisc;
1183 	netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1184 	if (dev_ingress_queue(dev))
1185 		dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1186 
1187 	timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1188 }
1189 
1190 static void shutdown_scheduler_queue(struct net_device *dev,
1191 				     struct netdev_queue *dev_queue,
1192 				     void *_qdisc_default)
1193 {
1194 	struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1195 	struct Qdisc *qdisc_default = _qdisc_default;
1196 
1197 	if (qdisc) {
1198 		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1199 		dev_queue->qdisc_sleeping = qdisc_default;
1200 
1201 		qdisc_destroy(qdisc);
1202 	}
1203 }
1204 
1205 void dev_shutdown(struct net_device *dev)
1206 {
1207 	netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1208 	if (dev_ingress_queue(dev))
1209 		shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1210 	qdisc_destroy(dev->qdisc);
1211 	dev->qdisc = &noop_qdisc;
1212 
1213 	WARN_ON(timer_pending(&dev->watchdog_timer));
1214 }
1215 
1216 void psched_ratecfg_precompute(struct psched_ratecfg *r,
1217 			       const struct tc_ratespec *conf,
1218 			       u64 rate64)
1219 {
1220 	memset(r, 0, sizeof(*r));
1221 	r->overhead = conf->overhead;
1222 	r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1223 	r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1224 	r->mult = 1;
1225 	/*
1226 	 * The deal here is to replace a divide by a reciprocal one
1227 	 * in fast path (a reciprocal divide is a multiply and a shift)
1228 	 *
1229 	 * Normal formula would be :
1230 	 *  time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1231 	 *
1232 	 * We compute mult/shift to use instead :
1233 	 *  time_in_ns = (len * mult) >> shift;
1234 	 *
1235 	 * We try to get the highest possible mult value for accuracy,
1236 	 * but have to make sure no overflows will ever happen.
1237 	 */
1238 	if (r->rate_bytes_ps > 0) {
1239 		u64 factor = NSEC_PER_SEC;
1240 
1241 		for (;;) {
1242 			r->mult = div64_u64(factor, r->rate_bytes_ps);
1243 			if (r->mult & (1U << 31) || factor & (1ULL << 63))
1244 				break;
1245 			factor <<= 1;
1246 			r->shift++;
1247 		}
1248 	}
1249 }
1250 EXPORT_SYMBOL(psched_ratecfg_precompute);
1251 
1252 static void mini_qdisc_rcu_func(struct rcu_head *head)
1253 {
1254 }
1255 
1256 void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1257 			  struct tcf_proto *tp_head)
1258 {
1259 	struct mini_Qdisc *miniq_old = rtnl_dereference(*miniqp->p_miniq);
1260 	struct mini_Qdisc *miniq;
1261 
1262 	if (!tp_head) {
1263 		RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1264 		/* Wait for flying RCU callback before it is freed. */
1265 		rcu_barrier_bh();
1266 		return;
1267 	}
1268 
1269 	miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1270 		&miniqp->miniq1 : &miniqp->miniq2;
1271 
1272 	/* We need to make sure that readers won't see the miniq
1273 	 * we are about to modify. So wait until previous call_rcu_bh callback
1274 	 * is done.
1275 	 */
1276 	rcu_barrier_bh();
1277 	miniq->filter_list = tp_head;
1278 	rcu_assign_pointer(*miniqp->p_miniq, miniq);
1279 
1280 	if (miniq_old)
1281 		/* This is counterpart of the rcu barriers above. We need to
1282 		 * block potential new user of miniq_old until all readers
1283 		 * are not seeing it.
1284 		 */
1285 		call_rcu_bh(&miniq_old->rcu, mini_qdisc_rcu_func);
1286 }
1287 EXPORT_SYMBOL(mini_qdisc_pair_swap);
1288 
1289 void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1290 			  struct mini_Qdisc __rcu **p_miniq)
1291 {
1292 	miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1293 	miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1294 	miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1295 	miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1296 	miniqp->p_miniq = p_miniq;
1297 }
1298 EXPORT_SYMBOL(mini_qdisc_pair_init);
1299